Docs / The method / Reading the results
Reading the results
Three views share one dataset: the dashboard for rates, the logs for individual verdicts, and the report for handing findings to whoever owns the fix.
The dashboard
The dashboard answers one question: what is running wrong, how often, and did it change. It is scoped to the latest completed evaluation day.
The rates, counts, trace identifiers and failure mode names used as examples throughout these docs are illustrative. They are not any customer's data, and they are not our own numbers either.
The tiles
| Tile | What it means |
|---|---|
| Error rate | The share of evaluated traces flagged by at least one monitored code on the latest completed day, with the change in points against the previous evaluation day. This is the lead number. |
| Traces evaluated | Traces scored on that day, with the all time figure underneath. This is your denominator, and it is the first thing to check when a rate looks dramatic. |
| Errors caught | Flagged traces on that day, and all time. |
| Codes monitored | Active monitored codes against the cap on your plan. |
Error rate by axial code
One line per monitored failure mode, plotted by trace date, so a spike lands on the day the traffic happened. Hovering reads a single day across every series, and clicking a legend chip mutes that line, which is the fastest way to see whether one code is carrying the whole aggregate.
Below the chart, one row per monitored code: its state as Active or Stopped, a sparkline of recent days, and its own rate with the underlying counts. Paused codes stay visible marked No longer sampled, so historical numbers remain readable without being mistaken for current ones.
The logs
Every individual verdict lands in Logs, filterable by code and by day. A row carries the trace, the verdict, the confidence and the evidence the judge relied on.
From a log row, Jump to highlighted evidence opens the trace at the passage the verdict rests on, with truncated and nested sections expanded around it. That path, from a rate on a chart to the words that justify one instance of it, is three clicks, and it is the check worth doing whenever a number surprises you.
The report
The report turns the same data into something you can send. It lists each monitored failure mode, how often it fired over the period, the counts behind each rate, and representative traces with their evidence, written so it can go to the team that owns the fix without a translation step.
Reading a number that moved
Before treating a change as a regression or a win, check these in order. Most surprising moves are explained by one of the first three.
- Check the denominatorA rate computed over eleven traces will swing violently. Read the counts under the rate, not just the percentage.
- Check trace day against evaluation dayRates are attributed to the day the traffic happened. A late or partial sync can make a recent day look anomalous simply because it is incomplete.
- Check whether the code changedIf the name or the scope changed, the series has a seam in it, and numbers either side of that seam are not comparable. See When a code changes.
- Check the traffic, not just the agentA new customer segment, a marketing push or a seasonal shift changes the input distribution. The agent can be unchanged and the rate still move.
- Then open three tracesRead the actual evidence for three flagged traces. If all three are convincing, believe the number. If one is wrong, the code needs re-scoping more than the agent needs fixing.
Traces where the material genuinely cannot decide the question are not negatives, and they stay out of the denominator. A code producing a lot of them usually needs its scope tightened rather than your agent needing a fix.
A cadence that works
- Daily, thirty seconds. Glance at the lead tile and the delta. Only dig in if something moved.
- Weekly, twenty minutes, together. Walk the per-code rows, open a handful of verdicts, and check whether any code has drifted away from what its name says.
- Per release. After shipping a targeted fix, read that code's line specifically, on trace dates after the deploy.
- Monthly. Re-read the taxonomy itself. Retire codes nobody acts on and free the slot.